Malignant skin cancer is one the most common and lethal type of skin cancer. Early detection of cancerous skin lesion will increase the possibility of patient survival. In recent years, implementation of models built on deep neural network in building medical diagnostic imaging system is quite beneficial to medical experts. In this study, we present an improved and fine-tuned EfficientNetB3 model to classify malignant skin lesion using concept of fine-tuning transfer learning. We have performed a comparative analysis of different deep learning pre-trained models like ResNet50, InceptionV3, InceptionResNetV2, EfficientNet B0-B2 models. The analysis findings signify the ability of utilizing fine-tuned EfficientNetB3 in the mission of melanoma detection and development of computer aided diagnostic system. All experimental results were carried out on ISBI-ISIC 2017 dataset. To check efficiency of proposed model, we compare the proposed model with EfficientNetB3 baseline model, present state-of-art pre-trained methods and approaches. The proposed EfficientNetB3 model obtained an accuracy of 87.12%, recall of 87.00%, precision of 87.00%, and F1 score of 85.00%. The proposed model achieved good computational results and also efficaciously addressed the problem of model overfitting and abate false negative labels.
Key words: Melanoma, Deep Learning, Classification, Skin Lesions, Computer Vision
|
